1 Gamma Two Cybernetic Brain

1.1 Overview

Where is my robot? You know - the one that acts like the ones in the movies; the one that I just tell what to do, and it goes out and does it. If it has problems, it overcomes them; if something in the world changes, it deals with the changes. The robot that we can trust to do the dirty, dangerous jobs out in the real world - where is that robot? What is preventing us from building and deploying robots like this? While there are a number of non-trivia and necessary hardware issues, the critical problem does not seem to be hardware related. We have many examples of small, simple systems that will (more or less) vacuum a floor, mow a lawn, or pick up discarded soda cans in an office. But these systems have a hard time dealing with new situations, like a tee shirt tossed on the floor or the neighbor’s cat sunning itself in the yard. We also have lots of tele-operated systems, from Predator aircraft to deep sea submersibles; from bomb disposal robots to remote controlled inspection systems. These systems can deal with changes to the world and significant obstacles: provided that one or more humans are in the loop to tell the robot what to do.

So what happens when a person takes over the joystick and looks through the low-resolution, narrow field of view camera of a perimeter-patrol security robot? Suddenly, where the robot was confounded by simple obstacles and easy to fix situations, the tele-operated system is able to achieve its goals and complete its mission. This is despite the fact that in place of a tight sensor-effector loop, we now have a long delay between taking an action and seeing the results (very long in the case of NASA’s Mars rovers). We have the same sensor data, we have the same effector capabilities, we have added a massive delay – yet the system performs better. If we had brains on our robots that could do the same task as the brains in living systems, the robots would act more like we expect, as opposed that provided by traditional robot control systems.

The work at Gamma Two is driven by the need to provide robots that can work with people. People have long and detailed memories coupled with amazing pattern recognition skills. This results in a significant challenge for robots that are intended to interact with people over long periods. People will quickly notice repetitive patterns, and expect new responses. At the same time, they expect 'predictable' behavior to increase trust, and enable the person to build a mental model of the robot. They also need the robot to be “smart enough” to provide a useful service, without being constantly monitored. In other words, they want an autonomous robot that is capable of performing useful tasks.

In order to understand why the Gamma Two Cybernetic brain can provide that functionality, we must briefly discuss the current state of the art in robotics. More detail on this can be found in the next section. Current work on robots can be roughly divided into two main types. These are:

1. Preprogrammed robots : These are robots with few sensors following pre-programmed scripts. They are generally used in industrial settings.

2. Autonomous robots: These are robots that do not follow a preprogrammed script. They generally choose between a set of possible behaviors.

Research on autonomous robots has generally been split into two camps.

1. Reactive systems: These are robots with a complement of specific behaviors triggered by specific sensor cues, or timing. They function well in the physical world because they react to changes in the environment. But they can not execute sophisticated plans, because they do not “know” what they are doing. This class includes the current generation of vacuum cleaning and lawn mowing robots

2. Deliberative systems: These are robots that can create complex and sophisticated plans. However, because these planning systems require symbolic inputs (and the world is not conveniently labeled), they can not be deployed in the physical world, but must be used in symbolic domains. These types of robots include chess playing robots and “soft” bots that operate on the Internet.

Living systems, humans included, interact with the world in a dynamic, and ever-changing way. Humans combine a rich reactive system with a sophisticated planning system. In order to do this, they take the take the data coming from their senses and turn this data into symbols, that can be used by the planning systems contained in the frontal lobes. We have built an analog to this translation ability in the Cybernetic Brain. This allows the Gamma Two Cybernetic Brain to achieve these goals, using a hybrid architecture based on the best models of how we want robots to behave – ourselves.

1.2 Classes of cognition deployed in robots

Before describing the architecture of the Cybernetic Brain, we will include a more detailed description of the traditional approaches to robotic control systems.

	1.2.1 Preprogrammed Robots This category covers more than industrial robots, but the best example is the robotic welder, seen on automotive assembly lines. It is typically a fixed base robot, which is programmed to execute a specific set of actions repetitively. A car slides into position, the arm swings into play, and welds a number of precise locations on the car body. These robots typically have low sensor density, in many cases no environmental sensors at all, and rely on a very tightly controlled environment for successful operation. The robots have very precise feedback on position and force, and rely on preprogrammed path planning to move efficiently from one position to another. Since the robots require a tightly controlled, deterministic environment, extreme steps are taken to remove both uncertainty and dynamic events from the workcell. These robots are deployed in factory settings, shipyards, laboratories, and large data warehouses. The industrial robot can be viewed as hardware controlled by low-level controller, which in turn are commanded by a pre-programmed sequence of instructions. Since there are few sensors, the instruction sequence is typically interrupted only if there is a safety violation, or an external command. The addition of safety monitoring was actually fairly late in the development, triggered by increasing numbers of injuries and deaths caused by the robots.
	1.2.2 Reactive Autonomous Robots Possibly the best known example of this type of robot is the Roombatm, from iRobot1. The reactive system relies on a higher density of sensors coupled with a behavior based control schema. In effect there is a relatively small set of different behaviors, triggered by direct sensor data. For example, the Roomba, when started in the center of a room, begins an expanding spiral pattern. This pattern continues until an obstacle is detected (by bump, boundary, or other sensors) at which point the spiral pattern is replaced by a random walk. The robot continues cleaning in straight lines, interrupted by obstacles for a period of time, and then switches to an edge cleaning behavior. If the Roomba is turned off, returned to the room center, and started again, the behavior will be identical, except for changes due to slightly different perceptions of the obstacles, or changes to the environment. In addition, if the Roomba gets stuck under the couch once, it will continue to repeat the actions resulting in getting stuck, over and over again. In the reactive model, the instruction sequence is replaced with a set of behaviors, and feedback is provided from the hardware up to the reactive controller. This sensor feedback is used (in conjunction with limited state data, such as elapsed time) to select between the available behaviors, The selection is primarily based on direct sensor input, so that the robot responds 'appropriately' based on the current configuration of the world. One drawback is that these robots frequently perform exactly the same sequence of actions in similar settings, with no regard to previous behavior. For a vacuum cleaner, not a big problem, but for the many reactive toys, it quickly becomes repetitive and predictable.
	1.2.3 Autonomous Deliberative Systems A deliberative system is one which maintains a representational model of the world, and selects sequence of actions intended to change the world in a deliberate way. These systems embody the traditional concept of artificial intelligence, and predominate in domains such as computer chess programs, on-line game playing software, and a few robots. These systems use sensor data to build a symbolic model of the world, and given a goal, generate an solution that would change the state on the world into the goal state. Most of these systems, require large amounts of computer power, and are still quite slow. When coupled with a robot, these problems become more complex. Now it is necessary to map the sensor based representation of the world into a symbolic representation that can be used by the deliberative system, and to map the symbolic solution into motor controls, and expected sensor states. This mapping must be done in real time, if the robot is to function in the real world. This has proven elusive; however, when done correctly, it results in the kind of robot that most people envision – one that 'knows' what it is doing, 'remembers' what it has already done, and 'overcomes' obstacles to its plan. However, it requires a complex solution, one that is a hybrid of the systems described above.

1.3 Structure of Gamma Two's Hybrid Architecture

The hybrid architecture that drives the Gamma Two Cybernetic Brain, is derived from living systems. The advantages are significant. Living systems have evolved to function in exactly the kind of domain we need robots to function in. These are a combination of very predictable interactions combined with uncertain and constantly changing situations. The tightly controlled, deterministic requirements of industrial robots fails because the world is constantly changing, and we cannot ask people to behave like machines. The reactive systems respond well to the dynamic environment, choosing behaviors deemed appropriate by their designers. But they will continue to do the same 'appropriate' behaviors over and over again, regardless of the responses of the people around them. Deliberative systems can react in an intentional way to changing situations, and can adapt their behaviors based on the responses, but they require large amounts of processing power, and are frequently too slow for use in robots.

The Gamma Two Cybernetic Brain integrates these three types of control structures in much the same way that mammalian brains integrate the biological analogs. Imagine that you are sitting at the breakfast table, finishing your morning tea. You are not consciously aware of the complex interplay of muscles, vision, and hearing that enable you to place the cup back onto the table, without spilling. This is a subtask that has been reduced to procedural memory – effectively a pre-programmed operation similar to one performed by an industrial robot. These types of 'muscle memory' actions are handled by the Perception/Action component of the Cybernetic Brain. When you decide it is time to get up from the table and leave for work, a complex process is executed in your pre-frontal cortex, that imagines the necessary actions that will result in achieving the goal of getting to work on time. This process takes into account the fact that this morning you will need to stop for gasoline (Semantic memory), and that yesterday there was road construction on your usual route (Episodic memory), so you decide to take an alternate. This is the function performed by the deliberative system in the Cybernetic Brain. As you are driving to work, you are constantly watching the behavior of the other cars on the road, and as they slow, you adjust you speed in response, when shifting lanes, you automatically check visually for an opportunity, adjust speed, and change lanes, in response to the data from your vision. This type of reactive behavior is handled by specific components in the Executive and Perception/Action Systems which perform complex sensor driven tasks with little or no involvement of the deliberative systems of the Cybernetic Brain.

By relying on the most effective technologies for the types of subtasks undertaken, the mammalian brain exhibits the kind of complex, thoughtful, yet responsive behavior that has eluded artificial intelligence researchers for decades. By designing and implementing a software system using biological principles, it is possible to achieve the same kinds of advantages in a robot, that living systems rely on.

1.4 Advantages of the Cybernetic Brain

If you look at the functional decomposition of the Cybernetic Brain, above, you might notice that there is no robot. Beyond providing an interaction pattern that is derived from living systems, the Cybernetic Brain is also intended to be nearly platform independent. Since it works with a mental model of the outside world, and a model of its own capabilities, the brain can be customized for almost any platform by changing the model of the platform. There are three components that must be customized

1. The Perception/Action Module;

2. The Reification Module; and

3. The Semantic Memory.

These components define the physical structure of the robot (the types and locations of the sensors), the mapping of raw sensor data to the symbolic representations of the data, and the knowledge of the world that is appropriate for the specific role of the robot. Part of this customization is the communication protocol between the brain and the body, what commands must be sent to move forward, grasp and hold an object, or make a facial expression. The brain worries about which expression is appropriate in the given situation, the body worries about how to move the motors and actuators to achieve the expression.

If the robot is already equipped with a low-level control system, we simply connect the brain to the board. If not, we design a specific micro-controller based component to act as a bridge.

1.5 Summary

Reactive systems are typically very sophisticated control mechanisms, based on either feedback control systems, reactive controllers, or behavior-based control mechanisms. These mechanisms are extremely powerful tools, provided that they are operating in tightly defined conditions. Unfortunately, the typical environment for an autonomous robot is anything but tightly defined. When robots are intended to work in conjunction with humans, the operating conditions are effectively uncontrolled, resulting in a number of common problems.

Deliberative systems are well developed mechanisms for the solution of hard symbolic problems. In a symbolic realm, such as chess, they can perform at the same level as the very best human brains. However, they are limited by only being able to process symbolic data. This means that they can not function in raw, unlabeled, human environments.

By merging the reactive and deliberative systems, the Gamma Two Cybernetic Brain has created a robotic system that can use the power of a deliberative planner and the robustness of a reactive system. This provides a significant improvement on both of the more traditional robot control mechanisms and answers the question “Where's my robot?” It is here, and it is real, now.